Osvaldo Pinali Doederlein reported 91 FPS on an AMD Radeon RX 9070 XT using Microsoft’s DirectX Raytracing sample with Clustered Geometry. The result came from a fixed-camera, high-geometry run at 1600 × 920, not a game benchmark; with the sample’s default geometry, he reported a much smaller gain.
What the RX 9070 XT test measured
In results posted on November 4, Doederlein compared conventional rebuilding of bottom-level acceleration structures (BLASes), which hold scene geometry, with Clustered Geometry in Microsoft’s D3D12RaytracingClusteredGeometry sample. His high-geometry run used 10,000 extra BLASes and a fixed camera. Doederlein’s test results
| Metric | Conventional BLAS rebuilding | Clustered Geometry | Test condition |
| Frame rate | 1.7 FPS | 91 FPS | 1600 × 920, 10,000 extra BLASes |
| Acceleration-structure rebuild time | 600 ms | 2.6 μs | High-geometry run |
| Ray-dispatch time | 8.5 ms | 8.2 ms | High-geometry run |
| Acceleration-structure memory | 7,518 MB | 771 MB | High-geometry run |
The frame-rate measurements work out to about 53.5 times the original rate, and the memory figures to roughly 90% less. Ray dispatch, the stage that sends rays through the scene, changed only slightly; the much larger differences were in rebuilding acceleration structures and their memory use.
Why geometry and resolution change the result
The reported gain depended heavily on the sample’s workload. With 100 extra BLASes, Doederlein reported about a twofold improvement; with 10,000, the frame-rate ratio exceeded 50 to 1. At 4K in the high-geometry run, the results were 1.6 FPS with conventional rebuilding and 28 FPS with Clustered Geometry, a 17.5-fold difference. He also reported that at 4K with less geometry, the two approaches could tie. Doederlein’s test results
That spread makes 91 FPS a figure for one demanding sample configuration, not a prediction for games. Doederlein cautioned that a real game does more work than the sample’s acceleration structures and ray operations. AMD likewise said games should not assume gains of this magnitude. AMD’s explanation of the preview features and sample results
What Clustered Geometry changes
AMD describes Clustered Geometry as breaking acceleration-structure builds into reusable clusters of up to 256 triangles. An application can assemble existing clusters rather than rebuild the structure from triangles. That targets the build work reflected in the sample’s large rebuild-time difference; it does not mean the ray-dispatch measurement itself rose by anything close to the frame-rate ratio.
The same preview SDK includes Partitioned TLAS, which focuses updates on changed regions of a scene. Indirect acceleration-structure builds take their arguments from GPU memory. A BLAS contains geometry; a top-level acceleration structure (TLAS) organizes instances of that geometry. These are tools for building and updating detailed ray-traced scenes, rather than a guarantee of a particular frame rate in a finished game. AMD’s feature descriptions
AMD’s listed preview hardware and setup
In its November 3 support announcement, AMD listed Radeon RX 9000 Series and Radeon AI PRO R9700 and R9600 Series GPUs for these preview DXR features. AMD’s requirements include Agility SDK 1.722-preview, AMD Software: Agility SDK Developer Preview Edition driver 26.20.12.01, DXC Shader Model 6.10 preview Patch 2 (v1.10.2605.37), Windows Developer Mode, and enabling D3D12ExperimentalShaderModels and D3D12RaytracingPreview. AMD’s support details
AMD distinguishes that preview setup from the retail Agility SDK 1.620 and driver 26.20.12.02. The listed preview DXR features require the Developer Preview driver, rather than the retail driver.
Microsoft’s separate RX 9070 result
AMD also reported a separate Microsoft sample test on a Radeon RX 9070, not Doederlein’s RX 9070 XT run. In that test, frame rate was 1.2 FPS with conventional rebuilding and 79 FPS with Clustered Geometry; per-frame acceleration-structure preparation and build time was 788 ms versus 3.11 ms. Total acceleration-structure memory, including TLAS, was 7,519 MiB versus 774 MiB. AMD described this as one test-playground configuration and said the system configuration was unspecified. AMD’s report of Microsoft’s test
